For future reference, it's a little easier to review renames if you make the rename a single commit and then subsequent edits to the file their own commit.

It looks like the two symbols this keeps in lib.npyio are BagObj and NpzFile. Both probably need to be public somewhere and np.lib.npyio is as good a place as any and has the benefit being where they publicly live now.

Given that this PR proposes keeping np.lib.npyio, It might make sense to deprecate (or just out-and-out move, since it's not used much) DataSource's location in the main namespace and move it to np.lib.npyio along with these two classes.

Can you add BagObj to the listing in doc/source/reference/routines.io.rst, along with NpzFile?

This removes 22 symbols from np.lib.__all__, all of them are in the main namespace:

In [5]: print(set(oldall) - set(np.lib.__all__))
{'save', 'polyint', 'poly1d', 'genfromtxt', 'polyadd', 'DataSource', 'savez_compressed', 'polyfit', 'savez', 'polymul', 'polysub', 'packbits', 'poly', 'roots', 'fromregex', 'loadtxt', 'load', 'polyval', 'unpackbits', 'polydiv', 'polyder', 'savetxt'}

In [6]: diff = set(oldall) - set(np.lib.__all__)

In [7]: diff.issubset(np.__all__)
Out[7]: True

Given that this PR proposes keeping np.lib.npyio, It might make sense to deprecate (or just out-and-out move, since it's not used much) DataSource's location in the main namespace and move it to np.lib.npyio along with these two classes.

Sure! DataSource is originally available in np.lib.npyio so I only removed it from the main namespace and added an expiration message - where to find it now.

Can you add BagObj to the listing in doc/source/reference/routines.io.rst, along with NpzFile?

NpzFile is already there, in the "NumPy binary files" section. I added BagObj to a new "Object wrappers" section.

nor can I find any interesting usage in the wild with code search right now

Hmm, for some reason I thought I saw some when I searched the other day but can't find any now. Sorry for the mistaken suggestion!

The only public usage I can find outside of numpy is here and it's not even used in that file.

Given this level of usage I agree with Ralf there's no reason to add it to the docs, it's probably safe to just delete it since it has no internal consumers. Maybe add a custom AttributeError message if someone tries to directly access it that encourages people to copy/paste the old implementation if they do need it, it's less than 20 lines of non-comment code.